You will be able to spot machine learning inside everyday apps and describe what each one learned from.
Think back over your morning. Your phone let you in by looking at your face. On the MRT you opened YouTube, and the first video was from a channel you watched once last week. At lunch you searched your photos for "receipt" to find a claim you forgot to submit, and the phone found it, even though you never tagged it. Then your bank sent an SMS asking you to confirm a payment you did make, at a shop you rarely visit.
None of those apps calls itself AI on the screen. All of them use machine learning, and once you can see it, you can ask better questions about every product that claims to be AI-powered.
Start with recommendations, such as the rows on Netflix, the next video on YouTube and the strip of suggested items under a product on Shopee, all of which learn from behaviour. Every click counts as an example, and so does every video you watch to the end and every item you add to your cart and then abandon.
They learn from more than just you. If many people who bought a rice cooker also bought a particular steamer tray, the system picks up that link and shows the tray to the next person who buys a rice cooker. That is why a feed can feel uncanny. It is matching you with people whose habits look like yours.
It also explains the familiar failure. Watch two cooking videos to help a friend and your feed fills with recipes for a week, because the system records the click and has no idea of the reason behind it. And because it learns from the past, it is slow to notice when your interests change.
Phone face recognition, photo search and voice typing are deep learning at work, the many-layered neural networks from lesson 1.1. These tasks are the fuzzy kind from lesson 1.2. Nobody can write rules for what makes your face your face in dim light with your glasses off, or what separates a receipt from a menu in a photo.
So these models were trained on very large numbers of images and recordings. A photo search model learned from images paired with descriptions of what is in them. A speech model learned from audio paired with the words that were said. Your phone runs the result.
The failures follow the training. Voice typing does well with the accents and words that were common in its training audio, and worse with the rest. If you have dictated a message with Singlish in it, or the name of a hawker centre, or a sentence that switches between English and Mandarin, you have probably watched it guess wrong. Face recognition can struggle in lighting it rarely saw. Those slips point straight back to what the model learned from.
Your bank's fraud alert and your email spam filter are both classifiers: models that sort each new case into one of a few groups, such as fraud or not fraud, spam or not spam. They are trained on past cases that people have already labelled. When a customer reports a transaction as fraud, or you mark an email as spam, that becomes a labelled example the system can learn from in future.
Inside, a classifier usually produces a score rather than a plain yes or no, such as how likely this payment is to be fraud, and the bank picks the cut-off that turns the score into an alert. Set it low and you get more alerts, and that SMS about a payment you really made is one of them. Set it high and genuine customers are bothered less, but more fraud slips through. A person at the bank decided where that line sits. The model only supplies the score.
"AI-powered" now appears on everything from aircons to CV screening software, and on its own the label tells you very little, since it could mean a deep learning model, a simple classifier, or a set of hand-written rules with new marketing.
Two questions cut through it. First, what did it learn from? A CV screening tool trained on a company's past hiring decisions will copy that company's past preferences, good and bad. A recommendation engine trained mostly on shoppers in another country may not suit people here. If the seller cannot tell you what the data was, that is an answer in itself.
Second, what does it do when it is unsure? A well-built system admits doubt. It passes the case to a person, asks you to confirm, or shows how confident it is. A weaker one guesses and gives no sign that it guessed. Your bank asking you to confirm a payment is the well-built version, even if it is mildly annoying at the checkout counter.
Lesson 2.3, A model can only be as good as its examples, looks at the data question in more depth. Until then, practise noticing where these models are hiding. Your phone is the easiest place to practise, because it is full of features that quietly learned from somebody's data, and the activity asks you to find a few of them.
Open three apps on your phone and write down one feature in each that likely uses machine learning, plus what data it probably learned from.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).